https://gcc.gnu.org/g:fada0de3559a530fa643a5de5fcee4f328699c0a

commit r17-6174-gfada0de3559a530fa643a5de5fcee4f328699c0a
Author: Kyrylo Tkachov <[email protected]>
Date:   Tue Nov 3 02:52:05 2026 -0800

    tree: Avoid building a zero tree for poly_int coefficients
    
    On targets with NUM_POLY_INT_COEFFS > 1 (AArch64, RISC-V), converting an
    INTEGER_CST to a poly_offset_int or poly_widest_int, as wi::to_poly_offset
    and wi::to_poly_widest do, fills the trailing coefficients using
    wi::ints_for<...>::zero.  For the tree-backed extended_tree types this
    called build_zero_cst, which looks up the INTEGER_CST cache on every
    conversion.
    
    These conversions are frequent.  For example get_ref_base_and_extent
    does them for every COMPONENT_REF, ARRAY_REF and MEM_REF it decomposes,
    so the alias oracle pays for it on each statement visited by the alias
    walkers.  In an AArch64 -O3 profile of a C++ file, wide_int_to_tree was
    about 1% of the compiler's instructions as self cost, mostly reached
    from get_ref_base_and_extent via refs_may_alias_p_2.
    
    extended_tree has a fixed precision independent of the type of the tree,
    so any zero INTEGER_CST represents the coefficient.  Return
    integer_zero_node instead.  The unextended_tree variant takes its
    precision from the type and is unchanged.
    
    The test input is the C++ test suite of fmtlib/fmt
    (https://github.com/fmtlib/fmt),
    built with CMake in Release mode (-O3 -DNDEBUG -std=gnu++11).
    
    instructions drop printf-test.cc -2.8%, format-test.cc
    -1.5%, all 8 translation units combined -1.9%.
    
    Bootstrapped and tested on aarch64-none-linux-gnu.
    
    gcc/ChangeLog:
    
            * tree.h (wi::ints_for<generic_wide_int<extended_tree<N>>,
            INL_CONST_PRECISION>::zero): Return integer_zero_node.
            (wi::ints_for<generic_wide_int<extended_tree<N>>,
            CONST_PRECISION>::zero): Likewise.
    
    Signed-off-by: Kyrylo Tkachov <[email protected]>

Diff:
---
 gcc/tree.h | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/gcc/tree.h b/gcc/tree.h
index c90b1bfe445a..4bdf87637949 100644
--- a/gcc/tree.h
+++ b/gcc/tree.h
@@ -7106,20 +7106,22 @@ wi::to_poly_wide (const_tree t)
   return t;
 }
 
+/* The precision of extended_tree does not depend on the type of the tree,
+   so any zero INTEGER_CST is a valid zero.  */
 template <int N>
 inline generic_wide_int <wi::extended_tree <N> >
 wi::ints_for <generic_wide_int <wi::extended_tree <N> >,
-             wi::INL_CONST_PRECISION>::zero (const extended &x)
+             wi::INL_CONST_PRECISION>::zero (const extended &)
 {
-  return build_zero_cst (TREE_TYPE (x.get_tree ()));
+  return integer_zero_node;
 }
 
 template <int N>
 inline generic_wide_int <wi::extended_tree <N> >
 wi::ints_for <generic_wide_int <wi::extended_tree <N> >,
-             wi::CONST_PRECISION>::zero (const extended &x)
+             wi::CONST_PRECISION>::zero (const extended &)
 {
-  return build_zero_cst (TREE_TYPE (x.get_tree ()));
+  return integer_zero_node;
 }
 
 inline generic_wide_int <wi::unextended_tree>

Reply via email to